Bayesian Optimization of Genetic Algorithm Hyperparameters in a Multi-Fidelity Framework for Efficient Lattice Material Design

This paper presents a three-level multi-fidelity framework for tuning genetic algorithm hyperparameters in lattice material design. A low-fidelity Gaussian process surrogate guides Bayesian optimization, a medium-fidelity 3D convolutional neural network rapidly evaluates properties, and high-fidelity FFT homogenization provides validation. Among the tested acquisition functions, logNEI performed best by modeling noise in genetic algorithm evaluations. The tuned configuration allowed a 25-generation GA run to reach an elastic modulus comparable to a full 75-generation optimization. Total computational cost fell from 225 to 171 hours, a reported 24% reduction, while a penalized objective reduced lattice evaluations with only minor performance loss.